| """Score a Laya checkpoint on the LocalLLaMA/typed-decisions test split. |
| |
| Metrics match the notebook's so the number is comparable to its published |
| 0.766, and to the Jev 1.13.0 reference row it quotes (0.727). |
| """ |
| import json, time, argparse |
| import numpy as np |
| from datasets import load_dataset |
| import laya |
| from laya.common import ece_score |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("model", nargs="?", default="out/laya-typed-decisions") |
| ap.add_argument("--dataset", default="LocalLLaMA/typed-decisions") |
| ap.add_argument("--config", default="all") |
| ap.add_argument("--split", default="test") |
| ap.add_argument("--device", default="cuda") |
| ap.add_argument("--by-workflow", action="store_true") |
| args = ap.parse_args() |
|
|
| ds = load_dataset(args.dataset, args.config, split=args.split) |
| agent = laya.Agent(args.model, device=args.device) |
|
|
| rows, lat = [], [] |
| for row in ds: |
| state, questions, gold = (json.loads(row[k]) for k in ("state", "questions", "gold")) |
| t = time.perf_counter() |
| res = agent.predict(state, questions) |
| lat.append((time.perf_counter() - t) * 1000) |
| rows.append((row.get("workflow", "?"), questions, gold, res["answers"])) |
|
|
|
|
| def score(subset): |
| acc, soft, brier, kl, tv, mae, w1, confs, corrects = ([] for _ in range(9)) |
| for _, questions, gold, pred in subset: |
| for qid, qdef in questions.items(): |
| p, g, t = pred[qid], gold[qid], qdef["type"] |
| if t == "choice": |
| keys = list(qdef["criteria"].keys()) |
| is_corr = float(p["choice"] == str(g["label"])) |
| pp = np.array([p["probabilities"].get(k, 1e-6) for k in keys]) |
| gp = np.array([g["probabilities"].get(k, 1e-6) for k in keys]) |
| elif t == "noul": |
| pv = p["noul"] |
| gv = g.get("noul", g.get("probabilities", {}).get("true", 0.5)) |
| is_corr = float(("true" if pv >= 0.5 else "false") == str(g["label"]).lower()) |
| pp, gp = np.array([1 - pv, pv]), np.array([1 - gv, gv]) |
| else: |
| n = len(qdef.get("criteria", [])) |
| pp = np.array([p["probabilities"].get(str(i), 0.0) for i in range(n)]) |
| gp = None |
| mae.append(abs(p["score"] - g.get("score", 0.0))) |
| w1.append(float(abs(p["score"] - g.get("score", 0.0)) <= 1.0)) |
| if pp.sum() > 0: |
| pp = pp / pp.sum() |
| p_lvl, conf = int(np.argmax(pp)), float(pp.max()) |
| else: |
| p_lvl, conf = int(round(p["score"])), 0.5 |
| is_corr = float(p_lvl == int(g.get("label", round(g.get("score", 0.0))))) |
| acc.append(is_corr); corrects.append(is_corr); confs.append(conf) |
| continue |
|
|
| pp = pp / pp.sum(); gp = gp / gp.sum() |
| acc.append(is_corr); corrects.append(is_corr); confs.append(float(pp.max())) |
| soft.append(float((pp * gp).sum())) |
| brier.append(float(((pp - gp) ** 2).sum())) |
| tv.append(float(0.5 * np.abs(pp - gp).sum())) |
| kl.append(float((gp * np.log(np.clip(gp / pp, 1e-12, 1e4))).sum())) |
| return dict(n=len(acc), acc=np.mean(acc), soft=np.mean(soft), brier=np.mean(brier), |
| kl=np.mean(kl), tv=np.mean(tv), |
| ece=ece_score(np.array(confs), np.array(corrects)), |
| mae=np.mean(mae) if mae else 0.0, within1=np.mean(w1) if w1 else 0.0) |
|
|
|
|
| m = score(rows) |
| print(f"\nmodel {args.model}") |
| print(f"cases {len(rows)} decisions {m['n']}") |
| print(f"accuracy {m['acc']:.3f} soft acc {m['soft']:.3f}") |
| print(f"ECE {m['ece']:.3f} Brier {m['brier']:.3f} KL {m['kl']:.3f} TV {m['tv']:.3f}") |
| print(f"score MAE {m['mae']:.3f} within-1 {m['within1']:.3f}") |
| print(f"latency p50 {np.percentile(lat,50):.1f} ms p95 {np.percentile(lat,95):.1f} ms per case") |
| print("\nreference (from the notebook, same test split):") |
| print(" TypeSafe Jev 1.13.0 accuracy 0.727 ECE 0.144 710 ms/case") |
| print(" notebook's 2xT4 finetune accuracy 0.766") |
| print(" teacher self-agreement accuracy 0.735 <- ceiling") |
|
|
| if args.by_workflow: |
| print() |
| for wf in sorted({r[0] for r in rows}): |
| s = score([r for r in rows if r[0] == wf]) |
| print(f" {wf:28s} acc {s['acc']:.3f} ECE {s['ece']:.3f} (n={s['n']})") |
|
|
